A statistical study on factors influencing piezoelectric properties of upside-down composites towards machine learning-driven development for recycling

• Incompatibility across models for specific compositing techniques is mitigated. • Different from conventional models, fitting constants become variables. • Contributors to upside-down composite functionality are established using new models. • Training dataset for artificial intelligence-driven, high-throughput recycling of piezoceramics is demonstrated. Using upside-down composites to recycle retired/discarded piezoceramics and then give them a second life in sensor applications paves the way towards sustainable production of piezoelectric materials. However, the piezoelectric properties of the recycled materials need to be significantly improved. The advancement of recycled materials should benefit from the recently developed AI-assisted methods, in order to minimize overconsumption of resources during the experimental trial and error. Previous works have identified obstacles to developing reliable models that can provide a systematic understanding of the contributors to the properties of the recycled materials. This work aims to overcome such obstacles by appropriately changing the fitting constants so that the models coincide with each experimental datapoint. These constants include descriptors of the microgeometry, size and orientation of the fillers, and the extent of polarization in the composites. By analyzing the variation of these constants between the datapoints, a clear perspective of the contributors to the properties of the recycled materials is established. This multi-variable approach is also extended to different fabrication techniques. The approach lays down a foundation for scaling up the optimization of the recycled materials by providing training and/or testing datasets for possible machine learning algorithms via potential high-throughput manufacturing routes.